Structured Generation

Pain Signals — You Need This Pattern When:

• LLM output is parsed with regex, string splitting, or "hope"
• Downstream code breaks when the LLM returns an unexpected format
• The same prompt sometimes returns JSON, sometimes prose, sometimes partial JSON
• You are writing brittle extraction code to pull structured data from free text
• Different calls to the same prompt produce outputs with different field names or shapes
• A parsing failure means the entire operation fails with no recovery path

Core Principle

An LLM is a function. Like any function, it should have a defined return type. Structured Generation constrains the output to a schema so the caller can rely on the shape without parsing or guessing.

The constraint happens at three possible levels — use the strongest one available: